Context Aware Task Orchestration With Deep Reinforcement Learning In Real Time Fog Computing Simulation Environment

Abstract

In the ever-evolving landscape of cloud computing, fog and edge computing have become more prominent because of their natural property of proximity to the demanding parts. Having all the heterogeneity on the processor side, task generators also have a variety of requirements in terms of complexity and latency. This dissertation gathered all of these different dimensions together in a robust ecosystem which can simulate a huge number of various scenarios that complies with the changing requirements for task orchestration and serves as a versatile platform for exploring task orchestration strategies. An advanced multi-layered cloud simulation model is proposed that intricately considers both low-level edge/fog and cloud environment constraints. At the core of the contribution lies a novel task orchestration model that transforms the orchestration process into reinforcement learning training steps. The proposed approach offers substantial advantages in terms of task succession, energy efficiency, and resource utilization. The system is evaluated with a custom developed simulation tool under different ambient fog computing conditions in terms of the edge device and the density of the task. The results of the experiment proved the superiority of the proposed system over the existing round-robin, heuristic-based, and PGOA algorithms with an overall increase in precision up to 28%. Besides, an efficient action space reduction technique is introduced to reduce the complexity of the action space, the proposed technique simplifies the decision-making process, leading to faster convergence and improved training efficiency.